Effiformer: a unified data-efficient vision transformer-CNN framework for interpretable epileptic seizure detection

Epileptic seizures are short episodes of abnormal electrical activity in the brain that can cause convulsions, loss of consciousness, and other similar symptoms. Despite therapy, around 30% of patients with epilepsy continue to have seizures, emphasizing the necessity for quick and effective detection measures. Accurate seizure detection allows for prompt intervention, which dramatically improves patient safety. This study has developed EffiFormer, a hybrid Vision Transformer-CNN model for reliable seizure detection using Electroencaphelogram (EEG) spectrograms. EffiFormer combines spatial feature extraction from EfficientNet with global attention mechanisms from the Data-Efficient Image Transformer, resulting in high accuracy even with limited training data. Our four-phase pipeline processes raw EEG signals through normalization, Short-Time Fourier Transform (STFT) to create spectrograms, synthetic seizure sample generation using SMOTE, and data augmentation. This study utilized a standard 60-20-20 training-validation-testing split for effectively training our proposed architecture and validating its accuracy. Further, this study tests the technique using the CHB-MIT dataset, which is a publicly available collection of long-term scalp EEG recordings from 22 pediatric patients at Children's Hospital Boston and covers a wide spectrum of seizure and non-seizure episodes. With an average sensitivity of 99.8% and average accuracy of 99.3%, our approach offers accurate seizure detection in various scenarios. Furthermore, explainable AI (XAI) approaches emphasize EEG regions that influence the model's results, making the model's decision-making process more transparent and interpretable for potential clinical review.